Discussion: SenseVoice/Paraformer models as backend option

## Discussion

VOSK is widely used for offline speech recognition. [SenseVoice](https://github.com/FunAudioLLM/SenseVoice) and [Paraformer](https://github.com/modelscope/FunASR) represent a newer generation of non-autoregressive models that could complement VOSK's model zoo.

### Comparison

| | VOSK (Kaldi-based) | SenseVoice | Paraformer |
|--|--|--|--|
| Architecture | HMM-DNN hybrid | Non-autoregressive Transformer | Non-autoregressive Transformer |
| Speed | Fast | 5x Whisper | 10x Whisper |
| Languages | Per-model | 50+ (single model) | Chinese/English |
| Model size | ~50MB per lang | 234M (all languages) | 220M |
| Accuracy | Good | State-of-the-art | State-of-the-art (Chinese) |
| Extra features | — | Emotion + audio events | Timestamps |

### Potential integration

SenseVoice/Paraformer models are available in multiple formats:
- ONNX (via Sherpa-ONNX, which already has a VOSK-like API design)
- GGUF (via FunASR-GGML)
- C++ (SenseVoice.cpp)

If VOSK supported these models as an additional backend, users would get significantly higher accuracy while maintaining the same offline, cross-platform deployment.

### Resources

- FunASR: https://github.com/modelscope/FunASR (16.6K stars)
- SenseVoice: https://github.com/FunAudioLLM/SenseVoice (8.3K stars)
- Sherpa-ONNX (similar API to VOSK): https://github.com/k2-fsa/sherpa-onnx